HF Model & Benchmark Matrix

24,078 variants · 14,984 scores · 120 benchmarks
Home/Models/google/gemma-4-31B-it

gemma-4-31B-it

google/gemma-4-31B-it

kind
original
quant
bf16 (16-bit)
family
gemma
modality
image-text-to-text
root model
google/gemma-4-31B
provider
google
license
apache-2.0
language
—
context
128K
library
transformers
vRAM (GB)
—
size on disk
—
downloads
12,401,314
likes
3,387
created
—
last modified
2026-07-20T16:42:18+00:00 (2026)
huggingface
https://huggingface.co/google/gemma-4-31B-it
tags
transformerssafetensorsgemma4image-text-to-textconversationalarxiv:2607.02770base_model:google/gemma-4-31Bbase_model:finetune:google/gemma-4-31Blicense:apache-2.0eval-resultsendpoints_compatibledeploy:azureregion:usdeploy:sagemaker

Benchmark scores (10)

benchmarkcategoryscorecontextconfevidence
AIME 2026math89.2AIME 2026 no toolsmedium| AIME 2026 no tools | 89.2% | 88.3% | 77.5% | 42.5% | 37.5% | 20.8% |
MATHmath85.6MATH-Visionmedium| MATH-Vision | 85.6% | 82.4% | 79.7% | 59.5% | 52.4% | 46.0% |
MMLU-Proknowledge85.2MMLU Promedium| MMLU Pro | 85.2% | 82.6% | 77.2% | 69.4% | 60.0% | 67.6% |
GPQA Diamondreasoning84.3GPQA Diamondmedium| GPQA Diamond | 84.3% | 82.3% | 78.8% | 58.6% | 43.4% | 42.4% |
LiveCodeBenchcoding80.0LiveCodeBench v6medium| LiveCodeBench v6 | 80.0% | 77.1% | 72.0% | 52.0% | 44.0% | 29.1% |
MMMUmultimodal76.9MMMU Promedium| MMMU Pro | 76.9% | 73.8% | 69.1% | 52.6% | 44.2% | 49.7% |
BigBenchreasoning74.4BigBench Extra Hardmedium| BigBench Extra Hard | 74.4% | 64.8% | 53.0% | 33.1% | 21.9% | 19.3% |
Humanity's Last Examknowledge26.5HLE with searchmedium| HLE with search | 26.5% | 17.2% | - | - | - | - |
Humanity's Last Examknowledge19.5HLE no toolsmedium| HLE no tools | 19.5% | 8.7% | 5.2% | - | - | - |
WER FLEURSspeech-asr8.0FLEURS (lower is better)medium| FLEURS (lower is better) | - | - | 0.069<sup>*</sup> | 0.08 | 0.09 | - |

Discussions (134)

titleauthorstatuscommentstype
When Gemma doesn't have the right tool schema it hallucinates it, create outputs of the fake-calls and use them as memory (WIP)luislozanog86open0disc
Chat template issue: extra newline character when rendering thinking?ggerganovopen0disc
Just when I spend 2 weeks researching how to improve natively Gemma's tools calls they fix them lolCoroneloopen0disc
Update READMEMaartenGrmerged0disc
Inconsistent Audio Support in Gemma-4-31B-IT: Video Audio Works, Standalone Audio Fails08mohitopen0disc
Raise per-image vision soft-token budget from 280 to 1120lucianommartinsopen0disc
Add response_template to tokenizer_config.jsonRocketknight1merged0disc
Rename README.md to README.amhgoosemaria149open0disc
Update README.mdPq234open0disc
gemma4-31bDraesondraft0disc
OpenPq234closed0disc
Add IFStruct v1.0 evaluation resultSaylorTwiftopen0disc
Add IFStruct v1.0 evaluation resultSaylorTwiftclosed0disc
Benchmarked on HexGrid Cloud : Gemma-4 31B + vLLM + RTX 6000 PRO : 1168 tokens/sec and still asking for more...hexgridcloudopen0disc
Interesting work — mobile AI perspective3morixdopen0disc
Add ask-my-agent bench v2 (Fallback) evaluation resultssapbotopen0disc
JobayerJy0018open0disc
Chat template may re-inject prior-turn reasoning during multi-turn tool use → repetition loopsManniX-ITAopen0disc
fix: chat template — null handling, reasoning preservation, turn-tag balance, input validationlucianommartinsmerged0disc
Love Gemma 4! A heartfelt plea from an independent developer regarding Antigravity CLI supportwalawalajeropen0disc
Add EvalEval community eval resultsEvalEvalBotclosed0disc
Chat template issues with multiple rounds of tool callingKimahrimanopen0disc
Repetition collapse when continuing an assistant turn (`continue_final_message=true`)CalinAstroBeamopen0disc
OpenCode Edits FailRiekeropen0disc
ReportFan-Puclosed0disc
26b (Dense ) + 31b (MoE) <=VS=> 26b (MoE) + 31b (Dense )Markobesopen0disc
Generating pictures with Gemma-4-31B-it (Try not to laugh)Flanuaopen0disc
Support OpenAI-compatible multimodal content part aliases in chat templateyasu-ohopen0disc
For Science (How model Thinking example)Flanuaopen0disc
Add MMMU-Pro evaluation resultSaylorTwiftmerged0disc
Fix empty thought channel in Gemma 4 chat templatedutifulbobclosed0disc
Tokenizer problems, or just quants?nesymerp1open0disc
Drops in performance randomly VLLM on H200 and B200 with 2 GPUsjtvinoopen0disc
gemma4-31b-it with MTP enabled works on dual DGX SPARKs nowkenleoopen0disc
Gemma4 is super slow on vllm docker , max to max 15 token per second only (4 X L4 GPU)vi-ajaybopen0disc
Add proper metadataosansevieromerged0disc
This model is so good, but the KV cache ruins it.UniversalLove333open0disc
Update README.mdhectorruiz9open0disc
o-O too good to be truemekasuopen0disc
could ya' add gpt j 6b to google ai studio? completion modeTralalabsopen0disc
Weird token after underscore `_//`yamikumodsopen0disc
CodeGemma 4 collectioncbrainopen0disc
fix(chat_template): Emit multimodal placeholders in tool response content-partsharshaljanjanimerged0disc
{Feature request] Eliminate pre-attention RMSNorm in Gemma 4 via scale invariance + weight foldinggraeficsopen0disc
Reviews of Gemma 4Juanoto2012open0disc
Update chat_template.jinja to address JSON Schema shapes that do not expose their meaning through a direct top-level typesigjhlopen0disc
Update chat_template.jinjasigjhlclosed0disc
Preserve anyOf/$ref/$defs in Gemma tool declarationssigjhlclosed0disc
Template fix: Preserve anyOf/$ref/$defs in Gemma tool declarationssigjhlclosed0disc
Possible chat_template.jinja issue: nullable $ref tool schemas are rendered as empty typessigjhlopen0disc
fix(chat_template): update SI and tool call handlingdougreidmerged0disc
JobayerJy0018draft0disc
When training models from the gemma4 series using GRPO, an abnormally high grad norm was observedmamazi00open0disc
add newlines and thinking tokens to template to avoid having to compute 3 extra tokens per generation in chat completion+reasoningquasar-of-mikusopen0disc
GOOLE WHERE IS MTP ?EvilinaMallerclosed0disc
Update README.mdhectorruiz9open0disc
Gemma 4 models are way to paranoid about dates, any tips?Ahugmopen0disc
Incorrect output in Gemma 4: seeking a solution to the problem ( la la la )Lintrariusopen0disc
Fix chat_template: emit empty <|channel>thought\n<channel|> wrapper for existing asst turnsflotherxiopen0disc
[Bug] chat_template: missing <|channel>thought\n<channel|> wrapper for non-thinking SFT / multi-turnflotherxiopen0disc
Thinking erratic at 30000+ contextJeslynMcKenzieopen0disc
Multilingual Support Listabcdvzzopen0disc
Will there be a small model like gemma-3-270m?ymckiopen0disc
Unexpected loss spikes and performance degradation when fine-tuning Gemma 4 (google/gemma-4-31B-it)rstaruchopen0disc
Add ParseBench evaluation resultsboyang-runllamaopen0disc
Will there be a small model for speculative decoding?Regrinopen0disc
Imagen 1 (2022) Should Be Open SourcedTralalabsopen0disc
Question about tool-calling order in chat_template.jinjajson0open0disc
gemma-4-31b-it unable to execute tool callingNaman2302open0disc
Do Gemma 4 models work well?Regrinopen0disc
fix: embed chat_template in tokenizer_config.jsonNERDDISCOopen0disc
Infinite loop is not fixed even with Google APIalexcardoopen0disc
Chat Template has a bug.Reithanopen0disc
why print rightarrowwangtf-Kevinopen0disc
Can anyone improve the model using the Rys methodology—by duplicating a block of layers?Regrinopen0disc
Update schemaRocketknight1merged0disc
Strange behaviour of the tokenizerandercorralopen0disc
Good Workflowanthoekfjopen0disc
fix: prefix-preserving tool calls + dialog compliancelucianommartinsmerged0disc
fix: function calling formatting in chat templateRyanMullinsopen0disc
fix-function-calling-chat-templateRyanMullinsclosed0disc
Chat template is too complicated that even Gemma 4 itself has no idea how to parse italexcardoopen0disc
Hardware requirementCharan01open0disc
Tokens per Image Parameter?buckeye17open0disc
Guys please add the MTP to this modelNarutoouzopen0disc
Will there be QAT models?Regrinopen0disc
Gemma 4 E4B will be as encyclopedically well-read as the 12b model?Regrinopen0disc
Create BTSdeletedopen0disc
brokersponsorBrokersponsoropen0disc
Update README.mdBrokersponsoropen0disc
Qusetion about math_vision and mmmu_pro evaluation resultJjjjjZzzopen0disc
The Gemma 4 model is great. But...suitup91open0disc
𝐎𝐍𝐋𝐘 𝐂𝐔𝐂𝐊𝐒 𝐀𝐍𝐃 𝐒𝐔𝐁𝐇𝐔𝐌𝐀𝐍𝐒 𝐅𝐎𝐋𝐋𝐎𝐖𝐒 𝐀𝚰 '𝐄𝐓𝐇𝚰𝐂𝐒' 𝐀𝐍𝐃 '𝐒𝐀𝐅𝐄𝐓𝐘'!Kickanopen0disc
junk outputsrirv938open0disc
file size so biga9secondsleeperopen0disc
How did you achieve such remarkable metrics?Regrinopen0disc
Update README.mddeveloperJenisclosed0disc
Rename README.md to README.md mmkkantromjjjopen0disc
token masking and space_between_special_tokens during finetuningDarkhnopen0disc
WHY IS NOT CODING WORK😤Drixpyopen0disc
Featherless AIMarcosFRGamesopen0disc
Fantastic Model for Legal/English Language Use Casesmd-1415open0disc
test it yourselfrosspanda0open0disc
It can't edit code!!deletedopen0disc
RuntimeError: FlashAttention forward only supports head dimension at most 256badhon1512open0disc
Can you guys make a smaller dense?nesymerp1open0disc
Missing 'reasoning' field in response when serving gemma-4-31B-it with vLLMIannopen0disc
Community resources for Gemma 4 deployment — mobile, local, and cloud pathsLinncharmopen0disc
tryingtiiihzxopen0disc
aavtiiihzxopen0disc
atiiihzxopen0disc
Unable to view the model file the architecture!naren-1219open0disc
The first small model for honest conversationghostwithahatopen0disc
Please, Google open-source the ancient model "Gemini 1.0 Pro". Please, devepoler comment and share with google.Tralalabsopen0disc
vllm docker version vllm/vllm-openai:gemma4 detect Qwen3ForCausalLM model architecture and only allow max_model_len 40960jiangtaozhclosed0disc
Fix Gemma4 `response_schema` regex to capture content after tool callsqgallouedecdraft0disc
..but where's the bigger one?jerryblunderopen0disc
This model didn't release, it escaped 💀🫪evewashereopen0disc
🚀测评视频+OpenClaw中使用方式leo009open0disc
Model seems to have issues in vLLM (characters duplication)dehnhaideopen0disc
Is a QAT fine-tune planned?stduhpfopen0disc
What is the training precision of Gemma 4? (bf16?)hyunmin253open0disc
Why Gemma4 can't recognize the entire text on image?alexcardoopen0disc
About the video understanding?Alexander-2open0disc
124B, pretty please?vody-amopen0disc
Sentencepiece model fileBingsuopen0disc
Are you insane to push agenda into local model?Krokozorclosed0disc
Any params/containers to deploy it from HF Inference Endpoints?fabriciocarraroopen0disc
Add AIME 2026 evaluation resultSaylorTwiftopen0disc
Add eval resultsmerveopen0disc
vllm / sglang support?mtclopen0disc
Best birthday gift ever.Darkknight535closed0disc
ordl-gemma-familyWinsocketclosed0disc
Proposed video configurationpcuenqmerged0disc
Hugging Face model & benchmark data scraped & matrix-ified. 24,078 variants · 14,984 scores · 120 benchmarks.